[PartialInliner] Skip call sites where inlining fails.

[fhahn]

Summary

InlineFunction can fail, for example when trying to inline vararg
fuctions. In those cases, we do not want to bump partial inlining
counters or set AnyInlined to true, because this could leave an unused
function hanging around.

[davidxl]

lib/Transforms/IPO/PartialInlining.cpp

933–934

Move the opt remark down.

[davide]

LGTM with David's suggestion of moving the remark down. Thanks!

[fhahn]

I've updated the patch to construct the remark before inlining, as the callsite gets erased by InlineFunction, but moved emission of the remark after the InlineFunction.

[fhahn]

lib/Transforms/IPO/PartialInlining.cpp

933–934

Done, thanks!

lib/Transforms/IPO/PartialInlining.cpp

for (User *User : Users) {
 
     if (IsLimitReached())
       continue;
 
     OptimizationRemarkEmitter ORE(CS.getCaller());
 
     if (!shouldPartialInline(CS, Cloner, WeightedRcost, ORE))
       continue;
 
-    ORE.emit([&]() {
-      return OptimizationRemark(DEBUG_TYPE, "PartiallyInlined",
-                                CS.getInstruction())
-             << ore::NV("Callee", Cloner.OrigFunc) << " partially inlined into "
-             << ore::NV("Caller", CS.getCaller());
-    });
+    // Construct remark before doing the inlining, as after successful inlining
+    // the callsite is removed.
+    OptimizationRemark OR(DEBUG_TYPE, "PartiallyInlined", CS.getInstruction());
+    OR << ore::NV("Callee", Cloner.OrigFunc) << " partially inlined into "
 
     InlineFunctionInfo IFI(nullptr, GetAssumptionCache, PSI);
-    InlineFunction(CS, IFI);
+    if (!InlineFunction(CS, IFI))
+      continue;
+
+    ORE.emit(OR);
 
     // Now update the entry count:
     if (CalleeEntryCountV && CallSiteToProfCountMap.count(User)) {
       uint64_t CallSiteCount = CallSiteToProfCountMap[User];
       CalleeEntryCountV -= std::min(CalleeEntryCountV, CallSiteCount);
     }
 
     AnyInline = true;

test/Transforms/CodeExtractor/PartialInlineNoInline.ll

+; RUN: opt < %s -partial-inliner -S -stats -pass-remarks=partial-inlining 2>&1 | FileCheck %s
+; RUN: opt < %s -passes=partial-inliner -S -stats -pass-remarks=partial-inlining 2>&1 | FileCheck %s
+
+@stat = external global i32, align 4
+
+define i32 @inline_fail(i32 %count, ...) {
+entry:
+  %vargs = alloca i8*, align 8
+  %vargs1 = bitcast i8** %vargs to i8*
+  call void @llvm.va_start(i8* %vargs1)
+  %stat1 = load i32, i32* @stat, align 4
+  %cmp = icmp slt i32 %stat1, 0
+  br i1 %cmp, label %bb2, label %bb1
+
+bb1:                                              ; preds = %entry
+  %vg1 = add nsw i32 %stat1, 1
+  store i32 %vg1, i32* @stat, align 4
+  %va1 = va_arg i8** %vargs, i32
+  call void @foo(i32 %count, i32 %va1) #2
+  br label %bb2
+
+bb2:                                              ; preds = %bb1, %entry
+  %res = phi i32 [ 1, %bb1 ], [ 0, %entry ]
+  call void @llvm.va_end(i8* %vargs1)
+  ret i32 %res
+}
+
+define i32 @caller(i32 %arg) {
+bb:
+  %res = tail call i32 (i32, ...) @inline_fail(i32 %arg, i32 %arg)
+  ret i32 %res
+}
+
+declare void @foo(i32, i32)
+declare void @llvm.va_start(i8*)
+declare void @llvm.va_end(i8*)
+
+; Check that no remarks have been emitted, inline_fail has not been partial
+; inlined, no code has been extracted and the partial-inlining counter
+; has not been incremented.
+
+; CHECK-NOT: remark
+; CHECK: tail call i32 (i32, ...) @inline_fail(i32 %arg, i32 %arg)
+; CHECK-NOT: inline_fail.1_bb1
+; CHECK-NOT: partial-inlining